Hi all,
This fixes the PR mentioned in the subject. When expanding
atomic_test_and_set we try the corresponding sync optabs and if none
exist, like for pre-SMP ARM architectures (-march=armv6 and earlier) the
code in optabs.c reverts to a single-threaded implementation that just
does a load and store.
However, if the result of the operation is to be ignored the code in
builtins.c follows the convention that the target RTX is set to
const0_rtx to signify that the result is ignored and the expand
functions in optabs.c are supposed to check for that and act
appropriately.
The code in the single-threaded codepath in expand_atomic_test_and_set
in optabs.c didn't perform this check and instead tried to emit a move
to a const0_rtx which ICEs further down the line.
This patch fixes that by checking if the result is ignored, and if it is
omits the load.
I wouldn't dare to remove the load in normal atomic handling code due to
all the memory ordering subtleties, but this code path occurs only when
we have a trivially single-threaded bare-metal target and the compiler
assumes no races anyway and no dodgy context switches and tries to
implement this with a ldrb+strb, so I think removing the ldrb is valid.
Bootstrapped on arm, aarch64 and x86 and tested as well.
Ok for trunk?
This appears on 4.9 as well, I'll test it on that branch as well
2014-07-16 Kyrylo Tkachov <kyrylo.tkachov@arm.com>
PR target/61713
* gcc/optabs.c (expand_atomic_test_and_set): Do not try to emit
move to subtarget in serial version if result is ignored.